<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>Programmer on mayulu的AI笔记</title><link>https://mayulu.co/tags/programmer/</link><description>Recent content in Programmer on mayulu的AI笔记</description><generator>Hugo</generator><language>zhcn</language><copyright>Copyright © 2025–2026 mayulu All Rights Reserved</copyright><lastBuildDate>Sat, 01 Aug 2026 10:48:06 +0800</lastBuildDate><atom:link href="https://mayulu.co/tags/programmer/index.xml" rel="self" type="application/rss+xml"/><item><title>Jeff Dean 提出的每个工程师都该懂的AI系统核心数字</title><link>https://mayulu.co/posts/latency-numbers-you-should-know-in-ai-age/</link><pubDate>Sat, 01 Aug 2026 09:33:19 +0800</pubDate><guid>https://mayulu.co/posts/latency-numbers-you-should-know-in-ai-age/</guid><description>Jeff Dean 经典的程序员延迟对照表，曾是传统分布式系统架构的性能参考基准。2026 年 Y Combinator 访谈中，他提出适配大模型时代的全新系统量化指标，将性能标尺从时间延迟转向带宽、运算能耗、芯片互联三大维度。数据搬运能耗可达计算的千倍，由此解释批处理、量化等工程优化的底层逻辑。文章对比新旧两套体系，梳理 LLM 真实推理时延，剖析专用推理硬件发展趋势，帮助 AI 开发者建立硬件量级直觉。</description></item></channel></rss>